嘘とワンダーランド

You know, I’m pretty lazy. I don’t want to do anything. I’d be happy to just read cracked.com and tvtropes.org all day. But I’m also a megalomaniac. I wanna know everything. And I mean everything. I wanna understand about the Great Vowel Shift, the colonization of Australia, the evolution of the influenza virus, the performance of text matching… everything. Unfortunately, for now at least, the human brain can’t know everything. You can’t just pour stuff in until you run out of it. Upload Wikipedia and be done with it. You need to actually work for it and study. Slowly, painfully and did I mention slowly?

Sure, you can use a SRS to manage information so that you don’t forget the stuff again. You only learn it once and then do your daily repetitions so it stays in your brain. That’s pretty cool already, but there are two drawbacks: it’s a little boring and, worse, tiring. It’s in fact so tiring that I had to give up an otherwise pretty neat sleeping schedule. I read a few articles and forum posts on this and everyone seems to agree – you can do about 200 repetitions per day, maybe 300. More than that is too tiring, too time-consuming, too boring and maybe just plain impossible. Sure, for a few days you can do more, but I never saw anyone maintain this. You just burn out. (But you can study more, just not via SRS.)

Of course, old repetitions and new facts compete for resources. You can’t just add and add facts. Soon your daily repetitions surpass 200 and you will forget them anyway. But what is the optimal course of action? Is it better to add facts slowly to minimize the risk of burn-out? Soon, university starts again and I’ll have to study for exams. That means adding facts I don’t really wanna learn that much (at least right now), but I’m also learning other things (2 languages, for example). So on one hand, I want to learn as comfortable as possible, but on the other hand, I need to get done in time, so I’d better be fast.

What do you do in such a situation? You run a simulation, of course! I expanded my handy SRS simulator and tested a few possible configurations. I’m just going to show you the 3 most interesting ones:

amount of daily repetitions for 200 days

new facts remaining each day

A few short notes first.

I set the amount of facts to learn to 3400. That’s my current number of new, unseen facts. Don’t ask how I got that many (*shame*, *shame*), but it’s not unusual for many learners. Of course, if you are learning basically forever, there won’t be a “last new fact” – you’ll constantly add new ones. In this case, just look at the first graph and find the plateau for each curve. You will never drop below this.

The 3 configurations are:

• A -> a maximum of 100 repetitions per day; add up to 50 new facts
• B -> a maximum of 200 repetitions per day; add up to 50 new facts
• G -> a maximum of 200 repetitions per day; add up to 200 new facts

The maximum is only respected when adding new facts, not when doing due facts.

Ok, now let’s have a look at the actual graphs.

The first and most obvious thing is that G totally fails at it’s goal. After about a week, the due cards explode right in your face and you’ll have to face up to 350 repetitions per day. And this continues to happen all the time, so you are pretty likely, at some point, to just give up. You feel so bad about the extreme and unexpected workload you’re facing that it becomes counter-productive. So to just “fill up” until you hit your maximum number of comfortable repetitions a day is a really bad idea.

Furthermore, we can see that both A and B are pretty good at keeping the maximum under control, A is pretty inefficient at it. You have lazy and busy days in a pretty regular pattern, but you never max out. Not really that good, but workable nonetheless. However, what you can see is that if you are working through a fixed amount of facts, B gets you into the “lazy phase” much sooner. A has many lazy days, but on average it is actually more work than B, not less.

The second graph shows us how fast we are progressing. It takes G about 50 days to work through all facts, B needs about 80 days and A about 150. Sure, G is pretty fast, but actually not much faster than B, at only about 40% less time. However, it is quite clear that A really is slow. It needs at least twice as long to work through the same amount of facts than the other two approaches.

Conclusion

I think the data allows us to make two important conclusions.

First, do not add facts like crazy. Trying to just add facts until you fall asleep might work for a few days, but very soon you hit the point where all those repetitions burn you out. They would demand up to twice the amount of work you were capable of just a week ago, so most likely, you will just fail to do them and forget everything again. Basically, you end up really tired, demotivated and not much smarter than before. A good waste of your time.

Second, to be lazy, work harder. It might sound counter-intuitive at first, but it makes sense. If you try to be lazy early on by working below your actual threshold, you will, on most days, actually have to work more than you expected. You will learn much slower and spend a lot more time on your repetitions. The reason for this is that you are spreading out your early repetitions, which are by far the hardest. If you work a bit more and get those early repetitions over with, it will get much easier later on. If you work consistently at your threshold, it wil be easier to make a routine out of it and your progress will be faster.

Finally, the simulation allows me to pick some more useful values. If you have another look at the second graph again, you can see that A and G look a bit like stairs, but B is smooth. This means there is an optimal value of new facts per day to pick that consistently maxes out your workload by looking at the slope of the curve. This optimal value, interestingly enough, is about 42. I do not believe this to be a coincidence.

If you start with a bad assumption, you will invariably reach bad conclusions and constantly delude yourself about that fact. The only way to fix this is to regularly question your own basic assumptions about things. The scientific method provides a neat way (in fact, the only way) to do just this. Ask yourself: Is there a different explanation? On what data do you base your decision? Does another interpretation fit, too?

I don’t just claim this, but actually live this way. I want to demonstrate this by changing my opinion about sleep. Frustrated with polyphasic sleep, I have reëvaluated my own assumptions and checked the data. I read more studies and biology texts, looked through my own records and re-read a few polyphasic blogs. And I must conclude that polyphasic sleep, by and large, doesn’t work.

Let’s start at the opposite end – what does work? Well, polyphasic sleep is the best (known) option you have when you can’t have more than 2-4 hours per day of sleep. If you must sleep that little, for example because you are into solo sailing or your newborn child and 2 jobs keep you up all day, than polyphasic sleep is right for you. It minimizes the damage this kind of life will do, but you will still be worse off. You will still be sleep deprived.

Ok, having acknowledged that, let’s start with the criticism. In fact, it’s a very simple criticism because it only involves one point.

Polyphasic sleep destroys your memory.

Sure, you are awake more (if you are lucky, most people aren’t and delude themselves to the fact), but you can’t use the time in any meaningful way. You can’t learn more, in fact, you’ll learn less. All existing studies show that performance is slightly below normal levels, which means you have 4-6 more hours of waketime, but you are actually performing worse than if you had slept them all. Great job. That’s like taking a shortcut, only to drive slower so that you arrive even later.

Why is there not a single polyphasic scientist? No, Tesla was not polyphasic, he crashed regularly. Edison lied about his schedule and, while being mostly polyphasic, didn’t save any time (and he was not a scientist). Buckminster Fuller only slept polyphasically when touring, for the reason I mentioned above.

Why is there not a single polyphasic polyglot? You’d think that someone who is learning multiple languages at the same time would be glad over every single hour per day he can get. Yet, not a single one of them is documented to be polyphasic. Some have tried (mostly early polyglots), no one was happy with it.

Why does no military or space agency advocate polyphasic sleep? There are several studies researching it, but they all document a severe loss of performance and they all advise against it, except when external circumstances force you to be polyphasic, as mentioned earlier.

Why does all data collected via SRS, like for example Supermemo, show that sleeping in big chunks correlates with good performance? If there are working examples of polyphasic sleepers, no one of them has ever demonstrated this via their SRS statistics, and Supermemo captures a lot of those. There isn’t a single example of someone sleeping 4 hours or less per day and still getting a normal retention rate for the same amount of data learned.

There is a simple answer to these questions: Because polyphasic sleep doesn’t work. It’s bullshit. For all the claims of “superhuman” feats, there hasn’t been a single bit of evidence for it. Proponents have made all kinds of claims and assurances, yet have presented nothing. Most of them don’t even seem to be capable of grasping the importance of empirical evidence. It is pseudoscience.

Conclusion

If you don’t care about your memory and you don’t care about being able to learn, sure, go right ahead. If you also keep in mind that the majority of people drop out of polyphasic sleep after a month or less, I would recommend a better alternative: Amphetamines. It has exactly the same amount of advantages (awake at all costs), is easier to use and fucks you up just the same.

I’m now recanting all my previous posts and claims about polyphasic sleep. They are wrong. I have marked the posts accordingly. Don’t sleep polyphasically, yo. If you still think that it works, prove it. And no, “I’ve been doing this for months and I’m fine!!1!” isn’t proof. Get some real data. A SRS is a good starting point. Show that learning for 8 hours or more per day works as least as well as normally and doesn’t destroy your sleep. You won’t be able to, but try all the same. Hint: just log your exact sleeping times and do a few standard performance tests. This alone will probably demonstrate that you are deluding yourself.

So what’s the real alternative? This.

There is a major philosophical schism in language study. Well, learning any skill in general. Let’s explore them via an example that I just today made a final decision on. If the following sounds to specific, then this is a misconception. It applies to any process of learning.

What’s the problem? Say you learn vocabulary via a SRS. That means adding the given item (be it character, word or sentence) as a two-sided card, with a question on the front and the answer on the back. In the case of Chinese characters, the question would be a keyword (or short phrase) containing the meaning of the character (e.g. “to guard” or “にち”) and the answer would be the actual character (e.g. “守” and “日”). You would see the meaning, remember the character and write it out (mentally or with a pen) and check the answer. So far, so good.

However, here comes the schism: Should you add hints to the question side, i.e. a picture, additional sound, any mnemonic you use or whatever kind of hints you can think of short of giving the answer away? There are two schools of thought here:

One group believes that you should study thoroughly and face hard goals, so that you will be able to get good results in the wild. In this case that means not giving any hint at all. You should have internalized these hints already and rely only on your memory. You wouldn’t have access to these hints in a real texts after all. The alleged advantage is a better retention rate, with the drawback of being slower and being slightly less fun. But in the end, it pays off. I will call this group the Nietzscheans because “what doesn’t kill me only makes me stronger.”

The other group thinks you should make the process of studying as easy and fast as possible. In the given example they would always give their mnemonic and maybe additional material on the question side. The alleged advantage is the ability to cover more material in the same amount of time and less exhaustion. Your real retention rate might be lower, but it will pay off anyway. I will call this group the LaFargueans because the exemplify the Right to be Lazy.

So, which group is right? Let’s make the example more concrete. I will be intensively studying a total of roughly 3500 characters so I can get up to speed on all recent Japanese literature. I learn about 100 new characters per day and do daily reviews, so this will take a significant of my time in the near future. So I am very much interested to decrease my workload.

I have read several reports and my own statistics to derive three basic approaches:

a) Use no hints. This make the card a bit harder to get right at first (80% retention rate), but after you have one good answer down, you will be fairly consistent (90% retention rate).
b) Use hints. The first review is somewhat easier (85% retention rate), but still involves some work to get it right. After this, you will almost always be right (95% retention rate).
c) Use hints at first, but drop them later. This is similar to b), but after 3 correct answers you switch off the hints. This will make you fail a lot more cards at first (70% retention rate), but you improve fast.

Now, how much reviews would you have to do daily with each approach? Graphs to the rescue!

daily reviews

same data, logarithmic scale

As you can see, using hints some or all of the time makes almost no difference with regards to reviews, but both approaches are quite a bit better than using no hints. How much better? Using no hints, the Nietzscheans will do about 47,000 reviews in 2 years. The LaFargueans will do about 37,000 reviews. That’s 10,000 reviews less, i.e. 20% less work! And because they are easier, they will also be faster. In my experience up to twice as fast.

How significant is this? Well, 20% less work means you can do 20% more in the same amount of time. The LaFargueans could learn 700 additional characters, not spending a minute more than the Nietzscheans. Of course they would argue that the LaFargueans won’t learn them as well. That’s true. Their real retention rate is lower. Let’s be generous and say the Nietzscheans have a retention rate of 90% after a year, while the LaFargueans have only 75%. But because the LaFargueans could cover a lot more material, they will know the same total of characters (3150), but their knowledge will be a lot more diverse. And they will be done faster (and have more fun). Instead of studying more, they could also stick with 3500 characters. That means they will have trouble with about 525 characters more. They could then spend their 20% more free time on them and would still be done faster (learning 525 cards separately would involve about 5000 repetitions, still 5000 less).

And this totally reflects my experience. Those that go the easier, more playful way always outperform the stricter ones. They can learn a lot more and see results earlier. And that’s why you should always cheat.

Let’s do some blatant propaganda for Spaced Repetition Software aka SRS.

Say, you want to learn something. Something big, like, Japanese or Chinese. Japanese uses 4 different writing system, but the one that stands out are the 漢字, i.e. the thousands of funny symbols. To be literate in Japanese, you need to now about 3000 of those. How would you learn something that huge?

To learn anything, you need two things. First, the information must be sticky. That means it must be represented in a form your brain can actually remember. What that means is: Ever tried remembering a long number? Like, 20 digits long? Impossible, unless you break it down. But ever remembered the whole plot, including all scenes, of a great movie? Totally easy. Your brain can remember pictures and narratives (related things, both by time and cause) easily, but abstract information is very hard. So you need to transform the 漢字, or whatever your learning, into pictures and stories, aka mnemonics. Fortunately, they were designed with that in mind, so that’s very simple.
Second, you need to review regularly. Your memory is leaky and needs constant reinforcment. Fortunately, every time the memory is refreshed, it will stick around a lot longer – roughly 2-3 times as long if you review just on the brink of forgetting. If you know some math, you’ll recognice this as an exponential progression. What does that mean? You only need to review about 7-8 times and the memory will stay for decades! So, that’s manageable. Unfortunately, the brain is a little faulty, so you will forget a few things anyway. The good thing is, though, that with very little effort, you can already reach a retention rate of 90-95%, so on average you only need around 10 reviews per fact to make sure you’ll remember it for a very long time.

That sounds pretty nice already, but still, 3000 漢字? Isn’t that a lot of work? No. That’s 3000 facts, meaning about 30,000 reviews. A review takes 10 seconds, at most. On average, it will take only about 5, but let’s assume 10. Worst case scenario, you know. In total, that’s only about 3.5 days of work. If it were not spaced out so much, you could finish it in a week. Sweet!

Have a look at those graphs.

3000 facts, 20 new facts a day

3000 facts, daily reviews

That’s your work over 10 months. The first shows how much reviews you will be doing per month in total. Yellow is the amount of new (or unseen) facts, red are reviews (or reps) of old facts. Below that is the amount of reviews per day for each month. As you can see, the daily workload is at most 20 minutes and goes does down rapidly. After 5 months, you know all 漢字 and will only be refreshing. And that’s only for a moderate amount of work with 20 new facts per day. You can easily do 50, or even 100 if you are determined. Pretty good, right?

Let’s look at what I’m doing right know. I’ve already learned around 2000 漢字 quite a while ago, but I’ve neglected a lot of them and I found I could only read them, not write them. Which sucks. So I started anew, but while I’m at it, I might as well do some more. :)

Chinese uses the 漢字 exclusively (almost), so you need more to be literate, around 4000 or so. A lot of those overlap (at least 60%) with Japanese, so it’s not a bad idea to learn the superset of 漢字 that both languages use. That would add up to about 5000-6000. I also want to read somewhat older literature and just love obscure sources, so I decided to totally rock the 漢字 and go for over 6000. A realistic upper bound is 8000 or so. After that you’ll have trouble finding any actual sources outside of taxonomy. I will also add 50 per day, on average. Right now, for the first 2000, I’m doing 100, but as they become more obscure, I’ll slow down to focus on more important aspects of the language. Still, 50 a day is maybe an hour of work. Let’s show some graphs.

8000 facts, 50 new facts a day

8000 facts, daily reviews

Less than an hour per day on reviews for half a year, then only up to 20 minutes. Half of them done in 2.5 months, meaning already pretty much 95% literacy. A total of 9 days of work for the reviews and 6 days for the initial learning. And that’s why I love my SRS.

For about 2 weeks I’ve been sleeping monophasically, trying a few different core durations. What I called “promising” back then, I call “waste of time” now. I see no significant change in my retention rate (95% for old facts, 84% for new facts[1]), but my motivation and ability to concentrate has dropped. I’m getting less done and can’t think straight.
I had hoped that being able to work for >6 hours at a time would improve my ability to program and read, but this isn’t the case. My log shows that the longest time I programmed in one sitting was roughly 1.5 hours. I think that “the flow” is overrated, or at least there are viable alternatives.

Anyway, I might as well go polyphasic again, so I’m back in the game. *sigh* I’d love to do uberman, but I know how life just loves to throw that one can’t-skip event in my way after 3 weeks or so. Sorry, if I had the money to make sure I could just tell everybody to fuck off for a few hours, sure. So it’s everyman again. Let’s hope I have enough discipline, experience and stuff to do to succeed this time. :)

Same bat schedule[2], same bat channel!

[1] Of course the exact number is more indicative of the specific structure and content of the facts, but given a workload of 20 new facts and ~80-100 reviews a day, it should be clear that my memory is working fine.
[2] 3h core at 20:00, naps at 5:00, 10:00, 15:00, optimized with daylight, 8 months of polyphasic experimentation and exams in mind